Muse Code Ships With Spark 1.2 and a $0.10 Tier

Claude Code, OpenAI's Codex and xAI's Grok Build had the terminal mostly to themselves, and on August 5, 2026 Meta shipped Muse Code into it in beta, running on a new model called Muse Spark 1.2. The prices did not come from Meta's launch post, which carries none: Hacker News commenter tristanj posted the developer-page figures, $1.25 per million input tokens and $4.25 per million output, or $0.10 and $0.20 for developers who opt in to letting Meta train on what they send.

Muse Code keeps its background agents alive for the whole session

Install is one line on macOS or Linux, curl -fsSL https://dev.meta.ai/install.sh | bash, per Meta's launch post. The design choice underneath is subagent persistence. Meta describes a simple main agent loop plus async background agents that "remain active throughout each session, rather than being spawned for individual tasks," gathering context once instead of repeatedly, deciding for themselves when to report back to the main agent.

The runtime is an append-only local event log. Every model call, tool run, approval and edit goes into it, which Meta says makes the agent "replay-exact and restart-safe": crash it, and it resumes exactly where it stopped. Muse Code ships with several default skills and the post details three, /plan (turns a task into an approval-gated plan), /grill (stress-tests that plan until it holds), and /goal (works toward a stated objective).

An append-only log that lets a crashed agent resume where it stopped is an unusual thing to build into a beta CLI.

Meta trained Spark 1.2 against the harness it ships inside

Muse Spark 1.2 is a coding-focused revision of Muse Spark 1.1, and Meta says it scaled up training compute on coding tasks while widening training-environment diversity. The unusual part is co-training: Meta trained the model on rejection-sampled trajectories from the Muse Code harness itself, plus "recipe optimizations for goals, compaction, and subagents" and integration of the Muse Code toolset. The model and the CLI were tuned against each other rather than shipped as separable products.

Meta also used Spark 1.1 to generate coding environments and instruction templates, then had it grade candidate solutions against those requirements, producing training data for 1.2. Its long-horizon claim rests on a kernel-optimization case study: over 1,000 tool calls, running up to 24 hours, writing, compiling and profiling Triton implementations of KDA and MLA kernels for NVIDIA Hopper GPUs, with third-party kernel libraries like FLA explicitly banned from the imports. The MLA run was benchmarked against a PyTorch reference at batch size 1, 64 heads, sequence length 8192 and latent dimension 512, with Spark 1.2 building a two-kernel pipeline that reuses the shared KV latent as both K and V.

Running a compile-profile-improve loop for 24 hours is a hard thing to stage for a launch post.

Meta will cut your bill tenfold if it can train on your code

The tiers, as posted by Hacker News commenter tristanj citing Meta's developer page:

Ten times off input, twenty times off output. Several commenters report the contributor tier is US-only, with the EU and Australia excluded. One developer who took Meta's $20 signup credits at the 1.1 launch also flagged new small print on those credits, "While using free credits your content may be used for product improvement," which they say was not there when the credits were issued.

The honest take: a published price on your own source code is more useful than the vague data-sharing toggles everywhere else, even if the number makes you flinch. OpenAI has run a version of this trade for a while, offering free API tokens in exchange for sharing feedback data, and the thread is full of people who could not work out whether they qualified. The commenter simonw put the trust question bluntly to a skeptic on Hacker News: "So you think the only difference between the $1.25/million token plan and the $0.10/million token plan is that you pay them more to both lie to you and breach their contractual obligation to you?"

Opus 5 beats Spark 1.2 on Meta's own benchmark chart

The top-voted critique in the Hacker News discussion (263 points, 164 comments) is about model selection. Meta benchmarked against OpenAI's mid-tier Terra rather than GPT-5.6 Sol and still dropped a benchmark to it, and left Claude Opus 5 in the chart where Spark 1.2 loses all but one. A later chart in Meta's post does include Sol, which lands above both.

Meta's own engineers are the other data point, and it is anecdotal. Multiple commenters claiming familiarity say Meta staff are still on Claude Code or Codex, with one writing that Meta "lets engineers use the best tools for the job" and nobody is rushing to switch. Publishing a chart where a competitor wins is at least honest framing, and choosing the opponent's second-best model to headline against undercuts it.

DeepSeek V4-Flash is the number the contributor tier was drawn against

Read the $0.10/$0.20 tier next to DeepSeek V4-Flash and the target is obvious, and the comparison runs through the top of the Hacker News thread. One commenter called it "a very interesting alternative to Deepseek for personal work where I don't care about the training," estimating similar cost per task with better benchmark numbers and built-in vision. Throughput reports on OpenRouter put Spark 1.2 around 150 to 180 tokens per second against roughly 50 for most DeepSeek providers, with time-to-first-token around 8 seconds.

DeepSeek keeps two advantages the discount does not touch. Weights and architecture get released back to the community, which nobody in the thread expects from Meta on this model. And third-party DeepSeek providers offer matching or lower pricing with zero-retention policies, so the cheap price and the data trade are not welded together the way they are here.

I would want those throughput numbers confirmed by someone other than an OpenRouter dashboard before betting a workflow on them.

Muse Code bills by the token with no ceiling you can set

Muse Code is built to run unattended for a long time. Meta's own case study has it making more than 1,000 tool calls across 24 hours, and the whole point of the restart-safe event log is that a crash does not end the run. Billing is per token, metered, uncapped.

Developers in the thread hit the wall immediately. Commenter sams99 entered a credit card and found no way to set a limit, only an email alert, and described it as being "one oopsie away from getting a 100 dollar bill." ray_kay777 said they would not try it at all without a top-up model. The workaround on offer is to route through OpenRouter, which enforces hard per-key limits, at the cost of the discounted first-party endpoint.

Muse Code will gate a file edit behind an approval prompt via /plan, and stress-test its own plan with /grill, and none of that machinery is wired to money. The empty slot is a budget the agent has to obey: a dollar ceiling handed to it at launch, enforced in the same runtime that logs every model call and tool run, refusing call number 1,001 when the money is gone.